TL;DR
This paper introduces HGNNPIP, a novel hybrid graph neural network that combines protein sequence and network information to predict protein-protein interactions (PPIs). It addresses the issue of high false positives in existing methods by proposing a new negative instance generation strategy and demonstrates superior performance across multiple species, including a case study in rice.
What Problem Does This Paper Address?
Understanding protein-protein interactions (PPIs) is crucial for deciphering biological functions, identifying drug targets, and designing novel therapeutics. While AI-driven computational methods have advanced PPI prediction, existing approaches often suffer from a significant problem: high false positive rates. This issue stems from how negative interaction examples (pairs of proteins that don't interact) are generated. Many methods rely on subcellular localization (SL) to create these negative examples, but this approach leads to a biased data distribution, where only a small fraction of proteins are considered for negative interactions. This bias results in models that might perform well on training data but generalize poorly to new, unseen data. The paper aims to overcome this limitation by developing a more robust prediction framework and a better strategy for generating negative interaction data.
Methods
The core of this work is the HGNNPIP (Hybrid Graph Neural Network framework for Protein-protein Interaction Prediction) model. HGNNPIP is a hybrid approach that leverages two key types of information:
- Sequence Embedding: It first converts amino acid sequences of proteins into numerical feature vectors using Word2vec for residue embedding, followed by a Deep Neural Network (DNN) to extract deeper, abstract features. This captures the intrinsic properties of individual proteins.
- Network Embedding: It then incorporates information about how proteins are connected in known PPI networks. This involves network partitioning (considering 1-hop and 2-hop neighbors), attention mechanisms to weigh the importance of neighbors, and topological feature extraction to represent the protein's role within the network.
These sequence and topological features are then concatenated and fed into a Multi-Layer Perceptron (MLP) classifier to predict the probability of interaction between a protein pair.
A significant methodological contribution is the proposal of three new strategies for negative instance generation:
- Popularity-biased Negative Sampling (PopNS): Samples negative pairs based on protein "popularity," defined by their degree (number of connections) in the PPI network.
- Similarity-biased Negative Sampling (SimNS): Selects negative instances based on the similarity between protein embedding vectors.
- Random Negative Sampling (RanNS): Uniformly samples negative instances from the network, aiming for an unbiased representation.
The paper argues that RanNS is the most effective strategy for reducing false positives and improving generalization.
Datasets
The study utilized six PPI datasets from various species for evaluation:
- S. cerevisiae (Yeast)
- C. elegans (Worm)
- E. coli (Bacteria)
- D. melanogaster (Fruit fly)
- Human
- O. sativa (Rice)
The first four datasets were sourced from existing work, while the Human PPI dataset was collected from another established source. A new PPI dataset for O. sativa was manually constructed from the STRING database. For evaluating the impact of data distribution bias, two testing sets ('NTS1' and 'NTS2') were created, each containing 100 randomly selected negative instances based on subcellular localization.
Key Insights and Ideas
- Subcellular localization is a flawed basis for negative sampling: The study strongly demonstrates that relying solely on subcellular localization for generating negative PPI examples introduces significant data distribution bias, leading to high false positive rates and poor model generalization.
- Random Negative Sampling (RanNS) is superior: Among the proposed negative sampling strategies, RanNS proved most effective in reducing false positives and improving overall model performance, suggesting its suitability for general PPI prediction.
- Feature fusion is critical: Combining protein sequence information with network topology features significantly boosts prediction performance, outperforming models that use only one type of information.
- HGNNPIP offers a good balance: The proposed HGNNPIP model achieves excellent performance with a relatively simple MLP classifier, indicating a good balance between predictive power and computational efficiency.
Key Results
- HGNNPIP outperformed four state-of-the-art (SOTA) models (DeepTrio, PIPR, DeepFE-PPI, and GAT) across six benchmark PPI datasets, showing superior accuracy, recall, MCC, and AUROC.
- The proposed RanNS strategy significantly reduced false positive rates on a challenging test set (NTS2) compared to traditional SL-based methods.
- The integrated HGNNPIP model, using both sequence and network features, showed improvements of 4.54-8.63% in Accuracy, 3.50-6.21% in AUROC, and 4.54-8.63% in AP compared to models using only sequence or network features.
- In a case study on Oryza sativa, HGNNPIP successfully predicted potential interacting partners of OsEDS1, and subsequent in vivo experiments validated the altered transcriptional levels of six predicted factors, suggesting their role in rice disease resistance.
Limitations and Future Work
- Protein structure integration: The authors acknowledge that incorporating protein structure information could further improve prediction performance.
- Parameter optimization: Parameter optimization was primarily performed on the S. cerevisiae dataset, and it's noted that these optimal parameters might not be universally applicable to all datasets.
- Comparison with other methods: While HGNNPIP was compared to several SOTA models, it remains unknown if it's superior to all other existing methods.
- Dynamics of interactions: The study focuses on static interactions; investigating the dynamics of protein-protein interactions would be a valuable future direction.
- Experimental validation: Future work includes generating transgenic plants with overexpressed or knocked-out target genes to further validate the functional roles of predicted interactions, particularly in the context of SCRE4-mediated host-pathogen interactions.
Summary updated